Okay now that we've looked at feeding logistics regression curves why don't we go ahead and see the output when we feed our Ravens data. So you feed the Ravens data with the GLM function. Here it operates just like the LN function the outcome is on the left hand side of a tilde and then the predictors on the right hand side. In this case the only difference is now we have to specify family equals binomial, and that's telling the GLM function that we have 01 data. If we had count data the bounded count data, real binomial data, then we would also have to give it a sample size. Also by default for the binomial and binary case, it's going to assume that the link function that you have is the logistical link function, which is what we want so that's fine. You see, the output when you do summary, summary works just like it did for Ellen. You see the output below. You ge the coefficients, in this case, -1.68 and for the Ravens coefficient you use 0.1. In this case on the Logit scale you want to look, for example, with the score variable, the Ravens score variable you want to see whether or not the coefficient is close to zero or not on the Logit scale. On the exponentiated scale you want to see whether or not it's close to one. That it'll give us the standard error and a Z value and the P value, we're going to interpret those just like we would in our linear model to acknowledging that this are drive in a different way. So here's the fitted curve, so this are the predictive, responses put back on the probability scale. So what R did is it plugged in the x values associated with the coefficient. So it took the scores, multiplied them times the coefficient of the scores, added the estimated intercept, and then took e to that value over 1 plus e to that value. And that gives us the probabilities. Now this is only part of the fitted S curve because the component of the S curve where the data actually are observed was restricted. So the S curve up here is one so the S curve looks something like this and then it would go down so you'll notice that this ends at 0.4 not zero. So that's the actual curve that's fitting, though here on the fitted values we're only showing part of it. If we were to exponentiate the ravens coefficients we get 0.1864 for the intercept and 1.1125 for the score. So that suggests an 11% increase in the probability of winning for every additional points that the Ravens score. Okay, that's how you would interpret this logistic regression coefficient. You can get confidence intervals for these two coefficients very easily with a confint operator. So on the output from the fitted model, we do confint, and then because most people, myself included, would prefer to look at these things on the exponentiated scale, you just exponentiate two endpoints with the x function. So what we see now is that our interval does contain one, it goes from 0.99 to 1.3. So it's saying that even though we for sure know that scoring points is what causes the Ravens to win the game, this coefficient turns out to not be significant, okay. The ANOVA function works just like it does in LM you put the output of the fitted model and you can put multiple models in there. For example nested models and ANOVA will give you a series of sequential tests. And so here, in this case, the variable is for the score, just the one variable that we're kind of interested in. So it just has a one Degree of Freedom test. This isn't that useful in this particular example, but if you had several models that you were interested in you'd put them into the ANOVA function. Or I think also it is especially useful if you have a factor variable, because sometimes you want the factor variable included or the factor variable, all levels of the factor variable removed. Then you might want to test whether or not that all levels of them are in total are necessary in the model. Notice that when you do summary and just get the coefficient table out from R, that test each one of the levels of the factor independently and doesn't test them all as a whole. So something like the ANOVA function is useful for putting a factor in and out of the model. So for interpreting our odds ratios remember that our odds ratios are not probabilities they're functions of probabilities but they're not probabilities. An odds ratio of one means no difference, okay? But on the Logit scale of the log odds ratio zero means no difference. In odds, [NOISE] often if we have an odds ratio below 0.5 or above two, that's considered, I would consider it a kind of strong effect. It depends very context dependent so if your working in the field of epidemiology or something like that, you often get very small odds ratios. 1.01 might be significant. You're looking at giant studies and the reason that these small odds ratios are important is because they're studying rather noisy things, like nutrition or something like that. How nutrition impacts health or something like that. And you wind up with because of all the various factors that incorporate, influence this study, you end up with very small odds ratios that are still meaningful, even if significant. On the other hand, you might run a very tightly controlled experimental clinical trial, or something like that, and then the odds ratios that you would want In order to declare something kind of meaningful, it would have to be much larger. So again, this is just less than 0.5 or bigger than two is a little bit of a benchmark. But remember, like all benchmarks, we only have a certain amount of utility. Really, how strong an odds ratio is relative to the scientific setting is incredibly dependent on the context that you're looking at in. So the relative risk is another entity that's often thought of in the same vein as the odds ratio, and many people like it. So the relative risk is just the ratio of the two probabilities. And many people like it because they tend to instinctively think a little bit better in terms of probabilities rather than odds. And relative probability seems like a reasonable thing to do. The problem with the Relative risk unlike the Odds is, it puts in some model constraints there quite hard, so we don't see relative risk regression for binary variable it's a very common thing to do. There are some software languages I know are has some packages to do it. It says has some packages to do it. One thing that is pretty common if you happen to have small probabilities then the relative risk approximates the odds ratios, this is an old very classic result. So that people go ahead and fit the Odds Ratios but then interpret them as if they're relative risks. So if you've seen that before then that's where that's coming from. So I give you couple more references here on Odds Ratios, but I think that's enough to get you started on generalizing your models for binomial data. I'm hoping that you can use all of your knowledge you got from linear models and help you work that into your use for generalized linear models. Next lecture, we're going to consider Poisson data. That's sort of the last real lecture of the series. Then we have kind of a some bonus fun content it's the last lecture, the very last lecture. So I look forward to seeing you next week. We're going to talk about poison random variables.
